Jim Chen (陳治閔), vice general manager of Delta Electronics.
Global digitalization is driving a surge in demand for AI data centers, which have become major sources of energy consumption and carbon emissions. With the global energy transition accelerating and companies under mounting pressure to meet net-zero goals, decarbonizing data centers has become an urgent task. This series, “Race to green data centers,” explores how businesses, technologies, and markets are shaping solutions and charting the path toward sustainable digital infrastructure.
As AI data centers drive up energy demand, the main challenge is not GPUs but securing power, said Jim Chen (陳治閔), vice general manager of Delta Electronics. In the U.S., it can take at least five years to obtain feeder line approvals, creating power shortage risks. In Taiwan, constraints come from limited renewable supply and a lack of baseload low-carbon energy. As a result, companies are expected to take on greater responsibility for their own power needs, with microgrids based on co-location expected to become a key area of development.
AI data centers prioritize energy independence, driving demand for microgrids
Demand for renewable energy among major corporations is growing more urgent. Beyond committing to RE100, many are now moving toward 24/7 low-carbon energy (CFE), which promotes round-the-clock use of zero-carbon power. More than 130 organizations worldwide, including Google and Microsoft, have joined the initiative.
“24/7 will reshape the entire grid structure,” said Chen. As data centers consume more energy, the challenge is not only access to green and low-carbon power but also to basis feeder supply. In the U.S., the three largest data center hubs—Texas, California, and Virginia—face an average feeder approval process of five years, with no guarantee of eventual connection.
According to DIGITIMES analyst Yu Pei-ju (余佩儒), the energy structure of AI data centers will gradually shift from centralized generation to distributed low-carbon sources, advancing toward a co-location model where data centers are sited alongside power assets. The ultimate goal is to reach 100% carbon-free energy while strengthening energy independence and operational flexibility.
The responsibility for securing electricity is gradually shifting from the government to enterprises. Yet not every company has expertise in power systems, which Chen said is where Delta holds an advantage. He explained that Delta has long worked on both on-site and off-site power equipment for data centers, with capabilities in both hardware and software. The company is now also using island-mode microgrids to address power shortages before facilities are connected to the grid.
Microgrids for data centers are designed to integrate diverse generation sources such as solar, hydrogen, and conventional generators, along with storage and management systems to optimize grid performance. As data centers place high demands on power quality, supply must remain stable and uninterrupted. Voltage and frequency fluctuations must be kept within 2%, supported by advanced control systems capable of forecasting electricity use.
Chen noted that power consumption at large data centers is uneven, rising and falling with AI-related demand. He emphasized that energy storage systems are the key component of microgrids, enabling integration of distributed generation sources while responding to rapid load fluctuations from AI workloads.
Chen observed that data centers are set to accelerate the deployment of microgrids. For example, if a data center’s client requires 100 MW of power, it must also secure 120 MW of backup capacity. At present, this backup is largely provided by diesel generators. In the U.S., hyperscale data center microgrids typically exceed 100 MW, with some reaching beyond 300 MW. Compared with microgrids built for commercial buildings or regional disaster resilience, those designed for data centers are on a much larger scale.
According to Chen, five projects are currently under discussion in the U.S. and could be implemented as early as next year. He added that the model is expected to be replicated in other locations between 2028 and 2029.
